New feature now available in Public Preview includes the ability to gain insight and receive troubleshooting documentation on failed environment builds, view and edit pipeline inputs & outputs before submitting a pipeline job, and speed up the checkpoint saving process and shorten the end-to-end training phase of large-scale distributed PyTorch models.
Troubleshoot common environment build errors - You can now quickly pinpoint warnings or errors in your environment definition with actionable troubleshooting recommendations to resolve your issue.
Submit pipeline jobs using the optimized submission wizard - You can now submit pipeline jobs through an optimized wizard experience. This new feature enables you to operate with larger screen space to enable pipeline inputs, change runtime settings, and decide whether or not to rerun the job before submitting the pipeline.
Streamline fast checkpointing with Nebula using ACPT - You can now use Nebula to help you to manage your checkpoint life cycles and enjoy a fast, asynchronous checkpoint saving experience in large scale jobs. This will ultimately shorten the end-to-end training time, and easily restore failed tasks – both greatly reducing expensive, unnecessary training costs.